A harness that can’t be stopped
isn’t a harness.
Vektra builds AI to keep people alive when the world breaks, and strong every other day. That only works if the people it serves can see it, steer it, and stop it.
The model proposes. The kernel decides. The human commands.
What we commit to.
Seven commitments. They hold for every division and every product, today and at scale.
Humans stay in command.
Consequential real-world actions wait for a person’s approval. A human stop is built to always work, and the system is designed to hold safe when it does.
Cognition is untrusted.
Language models are treated as untrusted. ILO is designed to place a small, auditable policy kernel between any model and the world.
Every action leaves a record.
An append-only audit journal with integrity checks, so any operator can reconstruct what happened and why.
Augment, never replace.
Nothing ships that makes the operator dumber, slower, or more dependent on the machine.
Honest status.
We say “in research” when it is in research. No capability claim ships without evidence.
Open for inspection.
Safety-critical code is meant to be read. We are preparing to release it so others can test it and break it.
Safe when degraded.
When signal, power, or the model drops, the system is built to hold, hand back, and say so.
What Vektra will not build.
Surveillance.
Trust is a decision filter. Products built to oppress disqualify themselves.
Autonomy without a stop.
No system that acts in the world without a human able to halt it.
Unearned claims.
No safety claim we cannot show.
What we study.
Research follows the hardest question in our mission: how a machine helps under pressure without taking control.
Agent oversight.
Approval ladders, policy gates, and stop conditions that hold even when the model is wrong.
Tamper-evident action logs.
Records an operator can trust after the fact, including when the system itself failed.
Degraded operation.
What an agent should do when the network, power, or model is lost.
Human-centered assistance.
Support that keeps operators clear and capable, including neurodivergent operators.
Where it stands.
Honest status. Research stays labeled until it is field-proven.
Intelligent Life Operator · In development
Policy kernel
Version 2.0, planned licence for the kernel
Founder-accountable.
Safety and security reports go straight to the founder. We answer honestly, fix what we can, and credit the people who found it.
Report an issue: security@vektraindustries.com (responsible disclosure). Research partnerships: contact@vektraindustries.com.
Pablo Navarro, Founder
